Fault Diagnosis of Pitch System of Wind Turbine based on Improved Stacked Auto-Encoder Network

Abstract

In order to improve the accuracy of fault diagnosis of the wind turbine's pitch system, an improved stack autoencoder network is proposed. Based on the Supervisory Control And Data Acquisition (SCADA) data of the wind turbine's electric pitch system, the batch normalization (BN) algorithm was introduced for the gradient dispersion problem in the feature extraction of ordinary autoencoder networks when there are many parameters. This article uses the Adam optimizer to iteratively update the neural network weights based on the training data. Then calculate the cross-entropy loss function and train the network with the minimum loss function as the goal. Finally, the Softmax classifier is used, and its output is the diagnosis and probability of each component of the pitch system. The data set in the pitch control system SCADA is selected. This paper selects the verification set in the pitch control system SCADA and substitutes it into the ordinary stack autoencoder and improved stack autoencoder network (SAE) for comparison and verification. The verification results show that the batch-standardized SAE network has a more optimized network model and higher recognition accuracy, and also provides a strategy for fault diagnosis of wind turbines.